Package org.apache.druid.segment

Interface DimensionMerger

All Known Subinterfaces:

DimensionMergerV9

All Known Implementing Classes:

AutoTypeColumnMerger, DictionaryEncodedColumnMerger, DoubleDimensionMergerV9, FloatDimensionMergerV9, LongDimensionMergerV9, NestedDataColumnMergerV4, NumericDimensionMergerV9, StringDimensionMergerV9

public interface DimensionMerger

Processing related interface A DimensionMerger is a per-dimension stateful object that encapsulates type-specific operations and data structures used during the segment merging process (i.e., work done by IndexMerger). This object is responsible for: - merging encoding dictionaries, if present - writing the merged column data and any merged indexing structures (e.g., dictionaries, bitmaps) to disk At a high level, the index merging process can be broken down into the following steps: - Merge segment's encoding dictionaries. These need to be merged across segments into a shared space of dictionary mappings: writeMergedValueDictionary(List). - Merge the rows across segments into a common sequence of rows. Done outside of scope of this interface, currently in IndexMergerV9. - After constructing the merged sequence of rows, process each individual row via processMergedRow(org.apache.druid.segment.ColumnValueSelector), potentially continuing updating the internal structures. - Write the value representation metadata (dictionary, bitmaps), the sequence of row values, and index structures to a merged segment: writeIndexes(java.util.List<java.nio.IntBuffer>) A class implementing this interface is expected to be highly stateful, updating its internal state as these functions are called.

Method Summary

Modifier and Type	Method	Description
ColumnValueSelector	convertSortedSegmentRowValuesToMergedRowValues(int segmentIndex, ColumnValueSelector source)	Creates a value selector, which converts values with per-segment, _sorted order_ (see DimensionIndexer.convertUnsortedValuesToSorted(org.apache.druid.segment.ColumnValueSelector)) encoding from the given selector to their equivalent representation in the merged set of rows.
boolean	hasOnlyNulls()	Returns true if this dimension has no data besides nulls.
void	processMergedRow(ColumnValueSelector selector)	Process a column value(s) (potentially multi-value) of a row from the given selector and update the DimensionMerger's internal state.
void	writeIndexes(List<IntBuffer> segmentRowNumConversions)	Internally construct any index structures relevant to this DimensionMerger.
void	writeMergedValueDictionary(List<IndexableAdapter> adapters)	Given a list of segment adapters: - Read _sorted order_ (e.

Method Detail

writeMergedValueDictionary

void writeMergedValueDictionary(List<IndexableAdapter> adapters)
                         throws IOException

Given a list of segment adapters: - Read _sorted order_ (e. g. see IncrementalIndexAdapter.getDimValueLookup(String)) dictionary encoding information from the adapters - Merge those sorted order dictionary into a one big sorted order dictionary and write this merged dictionary. The implementer should maintain knowledge of the "index number" of the adapters in the input list, i.e., the position of each adapter in the input list. This "index number" will be used to refer to specific segments later in convertSortedSegmentRowValuesToMergedRowValues(int, org.apache.druid.segment.ColumnValueSelector).

Parameters:

adapters - List of adapters to be merged.

Throws:

IOException

See Also:

DimensionIndexer.convertUnsortedValuesToSorted(org.apache.druid.segment.ColumnValueSelector)

convertSortedSegmentRowValuesToMergedRowValues

ColumnValueSelector convertSortedSegmentRowValuesToMergedRowValues(int segmentIndex,
                                                                   ColumnValueSelector source)

Creates a value selector, which converts values with per-segment, _sorted order_ (see DimensionIndexer.convertUnsortedValuesToSorted(org.apache.druid.segment.ColumnValueSelector)) encoding from the given selector to their equivalent representation in the merged set of rows. This method is used by the index merging process to build the merged sequence of rows. The implementing class is expected to use the merged value metadata constructed during writeMergedValueDictionary(List), if applicable. For example, an implementation of this function for a dictionary-encoded String column would convert the segment-specific, sorted order dictionary values within the row to the common merged dictionary values determined during writeMergedValueDictionary(List).

Parameters:

segmentIndex - indicates which segment the row originated from, in the order established in writeMergedValueDictionary(List)

source - the selector from which to take values to convert

Returns:

a selector with converted values

processMergedRow

void processMergedRow(ColumnValueSelector selector)
               throws IOException

Process a column value(s) (potentially multi-value) of a row from the given selector and update the DimensionMerger's internal state. After constructing a merged sequence of rows across segments, the index merging process will iterate through these rows and on each iteration, for each column, pass the column value selector to the corresponding DimensionMerger. This allows each DimensionMerger to build its internal view of the sequence of merged rows, to be written out to a segment later.

Throws:

IOException

writeIndexes

void writeIndexes(@Nullable
                  List<IntBuffer> segmentRowNumConversions)
           throws IOException

Internally construct any index structures relevant to this DimensionMerger. After receiving the sequence of merged rows via iterated processMergedRow(org.apache.druid.segment.ColumnValueSelector) calls, the DimensionMerger can now build any index structures it needs. For example, a dictionary encoded String implementation would create its bitmap indexes for the merged segment during this step. The index merger will provide a list of row number conversion IntBuffer objects. Each IntBuffer is associated with one of the segments being merged; the position of the IntBuffer in the list corresponds to the position of segment adapters within the input list of writeMergedValueDictionary(List). For example, suppose there are two segments A and B. Row 24 from segment A maps to row 99 in the merged sequence of rows, The IntBuffer for segment A would have a mapping of 24 -> 99.

Parameters:

segmentRowNumConversions - A list of row number conversion IntBuffer objects.

Throws:

IOException

hasOnlyNulls

boolean hasOnlyNulls()

Returns true if this dimension has no data besides nulls. See NullColumnPartSerde for how null-only columns are stored in the segment.